Training AI Agents to Code: The Playbook Everyone's Been Waiting For

Cobus Greyling put together a collection of patterns and best practices for what experts call "loop engineering"—basically, the art of designing systems where an AI doesn't just do one thing, it enters an orchestrated, intelligent cycle of prompting. Think of the recent work from Addy Osmani and Boris Cherny at Anthropic—that kind of thing, but documented and ready to use.

The problem is most people still treat AI agents like black boxes: throw in a request, cross your fingers. But the real power is in designing the right loops. How does it work? The AI produces a first output, you evaluate it, provide structured feedback, it iterates, and repeat—all in an orchestrated way, not haphazardly. It's like directing a rehearsal, not flipping a light switch.

This repository isn't abstract theory. It's packed with concrete patterns: how to structure prompts to get better outputs, how to organize feedback so the AI actually understands and uses it, how to avoid the usual pitfalls (infinite loops, hallucinations, non-convergence). Real examples, actual use cases, and the reasoning behind every choice.

Why does this matter right now? Because AI agents for coding are getting serious—they're not toys anymore, they're tools companies and developers can actually use. But if you don't know how to make them work properly, it's like having a Ferrari and driving it like a rental car. This playbook shows you how to push it to 300 mph.

The interesting part is that the inspiration source (Anthropic and their researchers) is becoming the reference point for this stuff. It's no accident the best patterns come from them—they've been playing with these systems for months and know what actually works. Greyling had the good sense to document it for everyone else.